Dont listen salesman talking about engineering. Unless he has credibility proven by things he had done recently. Person might be zillionare but he cant buy credibility.
The boundary is wherever you let go of determinism.
If you can deterministically turn specs into the same code every time, congratulations! But all you've done is made a higher level language.
LLMs are useful to bring more vagaries of the real world into the realm of determinism, and the tool we use to do that is code.
Specs fail to codify all kinds of externalities that become risks over time: how UIs look is subject to fashion trends, for example - and your users may be quite happy with the UI in the old version, but you regenerated it from specs and whoops! Now it's not accessible anymore! Also the OK button is now hooked up to a string that maps to an event handler that doesn't exist because it wasn't locked down in a spec. Whoops!
Hopefully you can see how it's not about code - what it's actually about is the way in which code codifies a desired result in a deterministic way, and does as efficiently as anything else we've invented so far.
Yes, in agentic coding, "code" is an artifact, in the sense of "release artifacts" the word used by VCS systems like GitHub. And the agentic step is not (or less) deterministic, so the burden on VCS is augmented by the need to track target code instead of being able to generate target code on demand through idempotency. Which explains why we are seeing the main VCS under infrastructural trouble.
Some recommendations to strengthen your theoretical framework and language.
Code is a quite ambiguous word, try using source code and target code. You can check out different definitions of source code, but a standard definition of these terms to me is Stallman's GPL, which is not just an academic textbook, it's actual binding contractual text for a huge body of software.
Compilers produce the same binary each time they are run on the same code in a given environment. Writing anything in response to anything tweeted by Elon is as useless as Elon Musk himself.
My 2 cents
Dont listen salesman talking about engineering. Unless he has credibility proven by things he had done recently. Person might be zillionare but he cant buy credibility.
What's stopping the salesman from lying about that too?
The experts aren't the ones buying. The expert and salesman are getting paid by the same person.
[dead]
The boundary isn't code vs specs.
The boundary is wherever you let go of determinism.
If you can deterministically turn specs into the same code every time, congratulations! But all you've done is made a higher level language.
LLMs are useful to bring more vagaries of the real world into the realm of determinism, and the tool we use to do that is code.
Specs fail to codify all kinds of externalities that become risks over time: how UIs look is subject to fashion trends, for example - and your users may be quite happy with the UI in the old version, but you regenerated it from specs and whoops! Now it's not accessible anymore! Also the OK button is now hooked up to a string that maps to an event handler that doesn't exist because it wasn't locked down in a spec. Whoops!
Hopefully you can see how it's not about code - what it's actually about is the way in which code codifies a desired result in a deterministic way, and does as efficiently as anything else we've invented so far.
Yes, in agentic coding, "code" is an artifact, in the sense of "release artifacts" the word used by VCS systems like GitHub. And the agentic step is not (or less) deterministic, so the burden on VCS is augmented by the need to track target code instead of being able to generate target code on demand through idempotency. Which explains why we are seeing the main VCS under infrastructural trouble.
Some recommendations to strengthen your theoretical framework and language.
Code is a quite ambiguous word, try using source code and target code. You can check out different definitions of source code, but a standard definition of these terms to me is Stallman's GPL, which is not just an academic textbook, it's actual binding contractual text for a huge body of software.
Compilers produce the same binary each time they are run on the same code in a given environment. Writing anything in response to anything tweeted by Elon is as useless as Elon Musk himself.